A Fortune 500 pharmaceutical

Clinical trial document summarization at scale

Healthcare 1-2 months by April 30, 2027
Share LinkedIn X

The problem

A pharmaceutical sponsor's medical directors read 100-page clinical trial protocols on a weekly cadence. Most of what they need is in 10 specific facts (inclusion/exclusion criteria, primary endpoints, safety endpoints, dosing, timeline, and 5 others). Build an agent that extracts these 10 facts from a protocol document, provides citations back to the source paragraphs, and flags discrepancies between similar trials in the same therapeutic area. Must handle PDFs. Must preserve traceability: no answer without a cited source paragraph.

What success looks like

Extraction accuracy above 85% on a 20-protocol test set. Every answer cites a source span. Discrepancy flags validated against sponsor's internal review.

Data access

Publicly available clinical trial protocols from ClinicalTrials.gov. Sponsor provides a labeled extraction schema.

Constraints

Must be reproducible. No hallucinated citations.

Reward

$5,000 team prize + sponsor team intro

Related tags

ragevaluationvision_language_models